fix: update current_turn_user_idx after preflight compression

[moizogg]

Problem

When preflight context compression replaces the messages list with a compressed [summary, ...tail] structure, the current_turn_user_idx becomes stale. This causes memory/plugin context injection to target the wrong message.

Impact

• Memory prefetch context may be injected into tool result messages instead of the user message
• Plugin context from pre_llm_call hooks may corrupt the summary message
• Can cause provider API errors from malformed role alternation
• Silent context loss in long conversations

Fix

After compression completes, scan the compressed messages list from the end to locate the current turn's user message and update both current_turn_user_idx and agent._persist_user_message_idx to the correct position.

[naqerl]

If the for loop would not found any messge, the original issue persists. It makes sense to set indexes to None in this case.

Nice to have these tests:

1. Mock compression and ensure that the index changed
2. Mock compression which deletes all user's messages and ensure, that stale index is not present

Also agent/file_safety.py and agent/context_references.py changes are unrelated and present in the another PR. It will cause merge conflict, so they could be safely removed

Thanks for your time!

[teknium1]

Thanks for identifying the stale-index hazard; the premise still holds on current main.

Problems

• The target code moved in 54870847cb0f530105907b1a793531b8d0f03d78: preflight compression now replaces messages in agent/turn_context.py:450, returns the old index at agent/turn_context.py:575, and agent/conversation_loop.py:801 still uses that index to choose the injection target.
• The scan in this PR leaves both indexes stale if no matching user message remains (agent/conversation_loop.py:690-692 on the PR head).
• Current main has additional message-replacing compaction paths at agent/conversation_loop.py:1034, 3143, 3398, 3621, and 4798; this hunk would not cover them.
• The diff has no regression coverage and includes unrelated logging edits in agent/context_references.py and agent/file_safety.py.

Suggested changes

• Move this into a shared rebinding helper, use it after each continuing compaction path, and clear both indexes when no current-turn user message survives.
• Add preflight and no-match regression tests, then split the unrelated logging edits.

Automated hermes-sweeper review.

[teknium1]

If the scan finds no matching user message, this leaves both indexes at their pre-compression values. Clear current_turn_user_idx and _persist_user_message_idx in that case so later injection and persistence cannot target a stale row.

[Fr4nZ82]

Production evidence that this and #43552 are the same stale current_turn_user_idx with two entry points (preflight compression here, message-alternation repair there): full timestamp forensics posted in #43552 (comment) — including a dropped 28 of 52 preflight cut followed 7s later by a request with no <memory-context> anywhere, while the next turn (list length unchanged) injected fine. A fix that re-derives the anchor right before the injection block would close both.

[teknium1]

Thanks @moizogg for the first report of this class. The preflight case is already covered on main: current_turn_user_idx is re-anchored at turn start. The remaining mid-turn gates (post-tool and pre-API compression) were fixed in #120175, merged as fd94981, which credits this PR. Closing as resolved on main.